A common computing architecture for processing relatively large amounts of data in a relatively short period of time includes multiple interconnected processors that share the processing burden. By sharing the processing burden, these multiple processors can often process the data more quickly than a single processor can for a given clock frequency. For example, each of the processors can process a respective portion of the data or execute a respective portion of a processing algorithm.
FIG. 1 is a schematic block diagram of a conventional computing machine 10 having a multi-processor architecture. The machine 10 includes a master processor 12 and coprocessors 141-14n, which communicate with each other and the master processor via a bus 16, an input port 18 for receiving raw data from a remote device (not shown in FIG. 1), and an output port 20 for providing processed data to the remote source. The machine 10 also includes a memory 22 for the master processor 12, respective memories 241-24n for the coprocessors 141-14n, and a memory 26 that the master processor and coprocessors share via the bus 16. The memory 22 serves as both a program and a working memory for the master processor 12, and each memory 241-24n serves as both a program and a working memory for a respective coprocessor 141-14n. The shared memory 26 allows the master processor 12 and the coprocessors 14 to transfer data among themselves, and from/to the remote device via the ports 18 and 20, respectively. The master processor 12 and the coprocessors 14 also receive a common clock signal that controls the speed at which the machine 10 processes the raw data.
In general, the computing machine 10 effectively divides the processing of raw data among the master processor 12 and the coprocessors 14. The remote source (not shown in FIG. 1) such as a sonar array loads the raw data via the port 18 into a section of the shared memory 26, which acts as a first-in-first-out (FIFO) buffer (not shown) for the raw data. The master processor 12 retrieves the raw data from the memory 26 via the bus 16, and then the master processor and the coprocessors 14 process the raw data, transferring data among themselves as necessary via the bus 16. The master processor 12 loads the processed data into another FIFO buffer (not shown) defined in the shared memory 26, and the remote source retrieves the processed data from this FIFO via the port 20.
In an example of operation, the computing machine 10 processes the raw data by sequentially performing n+1 respective operations on the raw data, where these operations together compose a processing algorithm such as a Fast Fourier Transform (FFT). More specifically, the machine 10 forms a data-processing pipeline from the master processor 12 and the coprocessors 14. For a given frequency of the clock signal, such a pipeline often allows the machine 10 to process the raw data faster than a machine having only a single processor.
After retrieving the raw data from the raw-data FIFO (not shown) in the memory 26, the master processor 12 performs a first operation, such as a trigonometric function, on the raw data. This operation yields a first result, which the processor 12 stores in a first-result FIFO (not shown) defined within the memory 26. Typically, the processor 12 executes a program stored in the memory 22, and performs the above-described actions under the control of the program. The processor 12 may also use the memory 22 as working memory to temporarily store data that the processor generates at intermediate intervals of the first operation.
Next, after retrieving the first result from the first-result FIFO (not shown) in the memory 26, the coprocessor 141 performs a second operation, such as a logarithmic function, on the first result. This second operation yields a second result, which the coprocessor 141 stores in a second-result FIFO (not shown) defined within the memory 26. Typically, the coprocessor 141 executes a program stored in the memory 241, and performs the above-described actions under the control of the program. The coprocessor 141 may also use the memory 241 as working memory to temporarily store data that the coprocessor generates at intermediate intervals of the second operation.
Then, the coprocessors 142-14n sequentially perform third—nth operations on the second—(n−1)th results in a manner similar to that discussed above for the coprocessor 141.
The nth operation, which is performed by the coprocessor 14n, yields the final result, i.e., the processed data. The coprocessor 14n loads the processed data into a processed-data FIFO (not shown) defined within the memory 26, and the remote device (not shown in FIG. 1) retrieves the processed data from this FIFO.
Because the master processor 12 and coprocessors 14 are simultaneously performing different operations of the processing algorithm, the computing machine 10 is often able to process the raw data faster than a computing machine having a single processor that sequentially performs the different operations. Specifically, the single processor cannot retrieve a new set of the raw data until it performs all n+1 operations on the previous set of raw data. But using the pipeline technique discussed above, the master processor 12 can retrieve a new set of raw data after performing only the first operation. Consequently, for a given clock frequency, this pipeline technique can increase the speed at which the machine 10 processes the raw data by a factor of approximately n+1 as compared to a single-processor machine (not shown in FIG. 1).
Alternatively, the computing machine 10 may process the raw data in parallel by simultaneously performing n+1 instances of a processing algorithm, such as an FFT, on the raw data. That is, if the algorithm includes n+1 sequential operations as described above in the previous example, then each of the master processor 12 and the coprocessors 14 sequentially perform all n+1 operations on respective sets of the raw data. Consequently, for a given clock frequency, this parallel-processing technique, like the above-described pipeline technique, can increase the speed at which the machine 10 processes the raw data by a factor of approximately n+1 as compared to a single-processor machine (not shown in FIG. 1).
Unfortunately, although the computing machine 10 can process data more quickly than a single-processor computing machine (not shown in FIG. 1), the data-processing speed of the machine 10 is often significantly less than the frequency of the processor clock. Specifically, the data-processing speed of the computing machine 10 is limited by the time that the master processor 12 and coprocessors 14 require to process data. For brevity, an example of this speed limitation is discussed in conjunction with the master processor 12, although it is understood that this discussion also applies to the coprocessors 14. As discussed above, the master processor 12 executes a program that controls the processor to manipulate data in a desired manner. This program includes a sequence of instructions that the processor 12 executes. Unfortunately, the processor 12 typically requires multiple clock cycles to execute a single instruction, and often must execute multiple instructions to process a single value of data. For example, suppose that the processor 12 is to multiply a first data value A (not shown) by a second data value B (not shown). During a first clock cycle, the processor 12 retrieves a multiply instruction from the memory 22. During second and third clock cycles, the processor 12 respectively retrieves A and B from the memory 26. During a fourth clock cycle, the processor 12 multiplies A and B, and, during a fifth clock cycle, stores the resulting product in the memory 22 or 26 or provides the resulting product to the remote device (not shown). This is a best-case scenario, because in many cases the processor 12 requires additional clock cycles for overhead tasks such as initializing and closing counters. Therefore, at best the processor 12 requires five clock cycles, or an average of 2.5 clock cycles per data value, to process A and B.
Consequently, the speed at which the computing machine 10 processes data is often significantly lower than the frequency of the clock that drives the master processor 12 and the coprocessors 14. For example, if the processor 12 is clocked at 1.0 Gigahertz (GHz) but requires an average of 2.5 clock cycles per data value, then the effective data-processing speed equals (1.0 GHz)/2.5=0.4 GHz. This effective data-processing speed is often characterized in units of operations per second. Therefore, in this example, for a clock speed of 1.0 GHz, the processor 12 would be rated with a data-processing speed of 0.4 Gigaoperations/second (Gops).
FIG. 2 is a block diagram of a hardwired data pipeline 30 that can typically process data faster than a processor can for a given clock frequency, and often at substantially the same rate at which the pipeline is clocked. The pipeline 30 includes operator circuits 321-32n, which each perform a respective operation on respective data without executing program instructions. That is, the desired operation is “burned in” to a circuit 32 such that it implements the operation automatically, without the need of program instructions. By eliminating the overhead associated with executing program instructions, the pipeline 30 can typically perform more operations per second than a processor can for a given clock frequency.
For example, the pipeline 30 can often solve the following equation faster than a processor can for a given clock frequency:Y(xk)=(5xk+3)2xk where xk represents a sequence of raw data values. In this example, the operator circuit 321 is a multiplier that calculates 5xk, the circuit 322 is an adder that calculates 5xk+3, and the circuit 32n (n=3) is a multiplier that calculates (5xk+3)2xk.
During a first clock cycle k=1, the circuit 321 receives data value x1 and multiplies it by 5 to generate 5x1.
During a second clock cycle k=2, the circuit 322 receives 5x1 from the circuit 321 and adds 3 to generate 5x1+3. Also, during the second clock cycle, the circuit 321 generates 5x2.
During a third clock cycle k=3, the circuit 323 receives 5x1+3 from the circuit 322 and multiplies by 2x1 (effectively left shifts 5x1+3 by x1) to generate the first result (5x1+3)2x1. Also during the third clock cycle, the circuit 321 generates 5x3 and the circuit 322 generates 5x2+3.
The pipeline 30 continues processing subsequent raw data values xk in this manner until all the raw data values are processed.
Consequently, a delay of two clock cycles after receiving a raw data value x1—this delay is often called the latency of the pipeline 30—the pipeline generates the result (5x1+3)2x1, and thereafter generates one result—e.g., (5x2+3)2x2, (5x3+3)2x3, . . . , 5xn+3)2xn—each clock cycle.
Disregarding the latency, the pipeline 30 thus has a data-processing speed equal to the clock speed. In comparison, assuming that the master processor 12 and coprocessors 14 (FIG. 1) have data-processing speeds that are 0.4 times the clock speed as in the above example, the pipeline 30 can process data 2.5 times faster than the computing machine 10 (FIG. 1) for a given clock speed.
Still referring to FIG. 2, a designer may choose to implement the pipeline 30 in a programmable logic IC (PLIC), such as a field-programmable gate array (FPGA), because a PLIC allows more design and modification flexibility than does an application specific IC (ASIC). To configure the hardwired connections within a PLIC, the designer merely sets interconnection-configuration registers disposed within the PLIC to predetermined binary states. The combination of all these binary states is often called “firmware.” Typically, the designer loads this firmware into a nonvolatile memory (not shown in FIG. 2) that is coupled to the PLIC. When one “turns on” the PLIC, it downloads the firmware from the memory into the interconnection-configuration registers. Therefore, to modify the functioning of the PLIC, the designer merely modifies the firmware and allows the PLIC to download the modified firmware into the interconnection-configuration registers. This ability to modify the PLIC by merely modifying the firmware is particularly useful during the prototyping stage and for upgrading the pipeline 30 “in the field”.
Unfortunately, the hardwired pipeline 30 may not be the best choice to execute algorithms that entail significant decision making, particularly nested decision making. A processor can typically execute a nested-decision-making instruction (e.g., a nested conditional instruction such as “if A, then do B, else if C, do D, . . . , else do n”) approximately as fast as it can execute an operational instruction (e.g., “A+B”) of comparable length. But although the pipeline 30 may be able to make a relatively simple decision (e.g., “A>B?”) efficiently, it typically cannot execute a nested decision (e.g., “if A, then do B, else if C, do D, . . . , else do n”) as efficiently as a processor can. One reason for this inefficiency is that the pipeline 30 may have-little on-board memory, and thus may need to access external working/program memory (not shown). And although one may be able to design the pipeline 30 to execute such a nested decision, the size and complexity of the required circuitry often makes such a design impractical, particularly where an algorithm includes multiple different nested decisions.
Consequently, processors are typically used in applications that require significant decision making, and hardwired pipelines are typically limited to “number crunching” applications that entail little or no decision making.
Furthermore, as discussed below, it is typically much easier for one to design/modify a processor-based computing machine, such as the computing machine 10 of FIG. 1, than it is to design/modify a hardwired pipeline such as the pipeline 30 of FIG. 2, particularly where the pipeline 30 includes multiple PLICs.
Computing components, such as processors and their peripherals (e.g., memory), typically include industry-standard communication interfaces that facilitate the interconnection of the components to form a processor-based computing machine.
Typically, a standard communication interface includes two layers: a physical layer and a services layer.
The physical layer includes the circuitry and the corresponding circuit interconnections that form the interface and the operating parameters of this circuitry. For example, the physical layer includes the pins that connect the component to a bus, the buffers that latch data received from the pins, and the drivers that drive signals onto the pins. The operating parameters include the acceptable voltage range of the data signals that the pins receive, the signal timing for writing and reading data, and the supported modes of operation (e.g., burst mode, page mode). Conventional physical layers include transistor-transistor logic (TTL) and RAMBUS.
The services layer includes the protocol by which a computing component transfers data. The protocol defines the format of the data and the manner in which the component sends and receives the formatted data. Conventional communication protocols include file-transfer protocol (FTP) and transmission control protocol/internet protocol (TCP/IP).
Consequently, because manufacturers and others typically design computing components having industry-standard communication interfaces, one can typically design the interface of such a component and interconnect it to other computing components with relatively little effort. This allows one to devote most of his time to designing the other portions of the computing machine, and to easily modify the machine by adding or removing components.
Designing a computing component that supports an industry-standard communication interface allows one to save design time by using an existing physical-layer design from a design library. This also insures that he/she can easily interface the component to off-the-shelf computing components.
And designing a computing machine using computing components that support a common industry-standard communication interface allows the designer to interconnect the components with little time and effort. Because the components support a common interface, the designer can interconnect them via a system bus with little design effort. And because the supported interface is an industry standard, one can easily modify the machine. For example, one can add different components and peripherals to the machine as the system design evolves, or can easily add/design next-generation components as the technology evolves. Furthermore, because the components support a common industry-standard services layer, one can incorporate into the computing machine's software an existing software module that implements the corresponding protocol. Therefore, one can interface the components with little effort because the interface design is essentially already in place, and thus can focus on designing the portions (e.g., software) of the machine that cause the machine to perform the desired function(s).
But unfortunately, there are no known industry-standard services layers for components, such as PLICs, used to form hardwired pipelines such as the pipeline 30 of FIG. 2.
Consequently, to design a pipeline having multiple PLICs, one typically spends a significant amount of time and exerts a significant effort designing “from scratch” and debugging the services layer of the communication interface between the PLICs. Typically, such an ad hoc services layer depends on the parameters of the data being transferred between the PLICs. Likewise, to design a pipeline that interfaces to a processor, one would have to spend a significant amount of time and exert a significant effort in designing and debugging the services layer of the communication interface between the pipeline and the processor.
Similarly, to modify such a pipeline by adding a PLIC to it, one typically spends a significant amount of time and exerts a significant effort designing and debugging the services layer of the communication interface between the added PLIC and the existing PLICs. Likewise, to modify a pipeline by adding a processor, or to modify a computing machine by adding a pipeline, one would have to spend a significant amount of time and exert a significant effort in designing and debugging the services layer of the communication interface between the pipeline and processor.
Consequently, referring to FIGS. 1 and 2, because of the difficulties in interfacing multiple PLICs and in interfacing a processor to a pipeline, one is often forced to make significant tradeoffs when designing a computing machine. For example, with a processor-based computing machine, one is forced to trade number-crunching speed and design/modification flexibility for complex decision-making ability. Conversely, with a hardwired pipeline-based computing machine, one is forced to trade complex-decision-making ability and design/modification flexibility for number-crunching speed. Furthermore, because of the difficulties in interfacing multiple PLICs, it is often impractical for one to design a pipeline-based machine having more than a few PLICs. As a result, a practical pipeline-based machine often has limited functionality. And because of the difficulties in interfacing a processor to a PLIC, it would be impractical to interface a processor to more than one PLIC. As a result, the benefits obtained by combining a processor and a pipeline would be minimal.
Therefore, a need has arisen for a new computing architecture that allows one to combine the decision-making ability of a processor-based machine with the number-crunching speed of a hardwired-pipeline-based machine.